Deep learning-based text error correction network model training method
Through the text error correction network model based on deep learning, the BERT language model is used for pre-training and fine-tuning, which solves the shortcomings of traditional text comparison systems in error correction accuracy and generalization capabilities, and achieves efficient and accurate correction of complex texts wrong.
Patent Information
- Application Number
- CN202411182355.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-27
- Publication Date
- 2024-12-03
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
When traditional text comparison systems deal with complex text error correction tasks, they cannot accurately capture semantic similarities or differences, and handle vocabulary diversity and grammatical structure differences, resulting in insufficient error correction accuracy and poor generalization ability.
Adopt deep learning-based text error correction network model, train error correction documents and corpus by obtaining and preprocessing, and use the BERT language model for pre-training and fine-tuning, to build a robust text error correction system, and capture professional vocabulary and self-created vocabulary in the field. Semantic understanding.
It realizes efficient and accurate error correction of text, improves error correction accuracy and generalization capabilities, and solves the problems of insufficient accuracy and poor generalization capabilities of traditional systems.
Smart Images

Figure CN119067101A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of text error correction, and in particular to a training method for a text error correction network model based on deep learning. Background Art
[0002] With the advent of the big data era, text data has exploded. Whether it is internal documents of enterprises, academic papers of scientific research teams, or tender information of bidding teams, they all contain a large amount of text information. The forms of these documents all reflect the professionalism and locality of the content; and the documents use a large number of professional vocabulary and self-created vocabulary in a certain field or direction, which greatly increases the difficulty of accurately understanding these text data;
[0003] When using traditional text comparison systems to handle complex text correction tasks, there are often limitations. They may not be able to accurately capture the semantic similarities or differences between texts, nor can they handle issues such as vocabulary diversity and differences in grammatical structures. As a result, traditional text comparison systems have insufficient text correction accuracy and poor generalization capabilities when handling text correction tasks. Summary of the invention
[0004] In view of the deficiencies in the prior art, the present invention provides a training method for a text error correction network model based on deep learning, which solves the problems existing in the prior art.
[0005] According to an embodiment of the present invention, a training method for a text error correction network model based on deep learning includes:
[0006] Obtain training error correction documents and several corpora;
[0007] Preprocess the training error correction documents and several corpora to obtain the preprocessed error correction task dataset and preprocessed corpus respectively;
[0008] Input the preprocessed corpus into the constructed BERT language model for training to obtain the BERT pre-trained language model;
[0009] The BERT pre-trained language model is fine-tuned according to the pre-processed error correction task dataset to obtain a text error correction network model.
[0010] As an embodiment of the present invention, the training error correction documents and several corpora are preprocessed to obtain a preprocessed error correction task data set and a preprocessed corpus, respectively, including:
[0011] Processing the obtained corpora according to a first preset processing strategy to obtain a preprocessed corpus; wherein the preprocessed corpus includes: a first training set, a first test set and a first valid set;
[0012] The acquired training error correction document is processed according to a second preset processing strategy to obtain a preprocessing error correction task data set; wherein the preprocessing error correction task data set includes: a second training set, a second test set and a second valid set.
[0013] As an embodiment of the present invention, the constructed BERT language model includes a plurality of encoders, and a plurality of encoder stacking settings, wherein the encoder includes:
[0014] A self-attention mechanism layer to capture the dependencies between positions in the input sequence;
[0015] The feedforward neural network layer is connected to the self-attention mechanism layer and is used to further process the output of the self-attention mechanism layer. The feedforward neural network layer includes two linear transformation layers and a nonlinear activation function. The first linear transformation layer maps the input vector to a high-dimensional space, and then passes through the activation function to increase nonlinearity. The output is mapped back to the original space or the required target dimension through the second linear transformation layer.
[0016] The normalization layer is connected to the self-attention mechanism layer and the feedforward neural network layer respectively, and is used to normalize the outputs of the self-attention mechanism layer and the feedforward neural network layer to a preset numerical range to accelerate the convergence speed of the model and improve the training stability.
[0017] As an embodiment of the present invention, the preprocessed corpus is input into the constructed BERT language model for training to obtain a BERT pre-trained language model, including:
[0018] Initialize the model parameters of the built BERT language model;
[0019] The first training set is input into the BERT speech model for iterative training, and after the training is completed, a training BERT language model is obtained;
[0020] Inputting the first test set and the first valid set into the training BERT language model, testing the training effect of the training BERT language model, obtaining the test results, and optimizing the model parameters of the BERT language model;
[0021] Repeat the above steps to obtain several test results and several trained BERT language models, and screen the several trained BERT language models based on the test results to obtain a BERT pre-trained language model.
[0022] As an embodiment of the present invention, the BERT pre-trained language model is fine-tuned according to the pre-processed error correction task data set to obtain a text error correction network model, including:
[0023] Load the BERT pre-trained language model;
[0024] The second training set is input into the BERT pre-trained language model for iterative training, and after the training is completed, a training text error correction model is obtained;
[0025] Inputting the second test set and the second valid set into the training text error correction model, evaluating the error correction performance of the training text error correction model, obtaining an evaluation result, and fine-tuning the model parameters of the training text error correction model;
[0026] Repeat the above steps to obtain several evaluation results and several training text error correction models, and screen the several training text error correction models based on the evaluation results to obtain a training text error correction network model.
[0027] A text comparison system, comprising:
[0028] A text collection module, used to obtain one or more error correction texts to be processed;
[0029] The text correction module is used to process one or more error correction texts to be processed using a pre-trained text correction network model to obtain the correct text after correction; wherein the text correction network model is obtained by the above-mentioned training method of the text correction network model based on deep learning.
[0030] A computer device comprising:
[0031] Memory for storing computer programs;
[0032] A processor is used to implement the training method of the text error correction network model based on deep learning as described above when executing the computer program.
[0033] A computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the training method of the text error correction network model based on deep learning as described above.
[0034] Compared with the prior art, the present invention has the following beneficial effects: by collecting a number of corpora and training error correction documents, with a robust configuration and sustainable training method, the problems of insufficient accuracy and poor generalization ability in the existing text comparison system are solved, and efficient and accurate error correction of the text is achieved. BRIEF DESCRIPTION OF THE DRAWINGS
[0035] Figure 1 A schematic diagram of a flow chart of an embodiment of the present invention;
[0036] Figure 2 A schematic diagram of preprocessing of a corpus according to another embodiment of the present invention;
[0037] Figure 3A schematic diagram of preprocessing of a training error correction document according to another embodiment of the present invention;
[0038] Figure 4 Schematic diagram of a BERT language model according to another embodiment of the present invention;
[0039] Figure 5 A self-attention score matrix calculation diagram of a self-attention mechanism layer according to another embodiment of the present invention;
[0040] Figure 6 A schematic diagram of the training process of a BERT pre-trained language model of a self-attention mechanism layer according to another embodiment of the present invention;
[0041] Figure 7 A schematic diagram of the training process of a text error correction network model of a self-attention mechanism layer according to another embodiment of the present invention. DETAILED DESCRIPTION
[0042] The technical solution of the present invention is further described below in conjunction with the accompanying drawings and embodiments.
[0043] like Figure 1 As shown, the embodiment of the present invention proposes a training method for a text error correction network model based on deep learning, including:
[0044] Obtain training error correction documents and several corpora;
[0045] Preprocess the training error correction documents and several corpora to obtain the preprocessed error correction task dataset and preprocessed corpus respectively;
[0046] Input the preprocessed corpus into the constructed BERT language model for training to obtain the BERT pre-trained language model;
[0047] Fine-tune the BERT pre-trained language model based on the pre-processed error correction task dataset to obtain a text error correction network model;
[0048] The working principle of the above technical solution is as follows: in actual use, by collecting a large-scale corpus and a large number of corpora from various sources; at the same time, preprocessing the training error correction documents to obtain a preprocessed error correction task data set, and obtaining a preprocessed error correction task data set and a preprocessed corpus; inputting the preprocessed corpus into the constructed BERT language model for training to obtain a BERT pre-trained language model; enabling the BERT pre-trained language model to accurately understand professional vocabulary and self-created vocabulary in a certain field or direction; then fine-tuning the BERT pre-trained language model according to the preprocessed error correction task data set to obtain a text error correction network model;
[0049] The beneficial effects of the above technical solution: Through the above technical solution, by collecting a number of corpora and training error correction documents, with a robust configuration and sustainable training method, the problems of insufficient accuracy and poor generalization ability existing in the existing text comparison system are solved, and efficient and accurate error correction of text is achieved.
[0050] In one embodiment, the training error correction documents and several corpora are preprocessed to obtain a preprocessed error correction task dataset and a preprocessed corpus, respectively, including:
[0051] Processing the obtained corpora according to a first preset processing strategy to obtain a preprocessed corpus; wherein the preprocessed corpus includes: a first training set, a first test set and a first valid set;
[0052] The acquired training error correction document is processed according to a second preset processing strategy to obtain a pre-processed error correction task data set; wherein the pre-processed error correction task data set includes: a second training set, a second test set, and a second valid set;
[0053] The working principle and beneficial effects of the above technical solution are as follows: Figures 2 to 3 As shown, when several corpora are processed according to the first preset processing strategy, the collected corpora are preprocessed by text cleaning, word segmentation, stop word removal and other preprocessing steps to obtain a pre-release corpus; for example, to prepare a data set in the pre-training process, a representative Chinese corpus is collected on the Internet, and the SIGHAN+Wang271K Chinese error correction data set is obtained by converting the original SIGHAN13, 14, and 15 data sets and the Wang271K data set format; the data of NLPCC 2018+HSK+CGED16, 17, and 18 are preprocessed by character segmentation, traditional Chinese to simplified Chinese, and data order disruption to generate familiar corpora for error correction; after processing these corpora, a preprocessed corpus is obtained, which is then divided into a first training set, a first test set, and a first valid set; the format of the generated data set is unified as id: unique identifier, original error text, subscript of the position of the error word in the sentence, and corrected text;
[0054] Prepare the error correction task dataset for the fine-tuning stage, that is, make the relevant text data into a dataset that conforms to the training set format according to your own needs. First, convert the collected error correction training documents into tsv files, that is, remove punctuation marks, sentence processing, and stop words from the text, and then use the Chinese NLP data enhancement preprocessing tool to perform data enhancement on the tsv format file, replace typos, synonyms, add words, etc. on the segmented text, and obtain the preprocessed error correction task dataset, which is then divided into the second training set, the second test set, and the second valid set;
[0055] In the process of dividing the preprocessed corpus and the preprocessed error correction task data set, the same division method is used, that is, 85% of the data is divided into a training set, and then the remaining data is randomly shuffled, and 35% of the remaining data is extracted as a test set. The valid set is the same as the test set.
[0056] In one embodiment, the constructed BERT language model includes a plurality of encoders, a plurality of encoder stacking settings, and the encoders include:
[0057] A self-attention mechanism layer to capture the dependencies between positions in the input sequence;
[0058] The feedforward neural network layer is connected to the self-attention mechanism layer and is used to further process the output of the self-attention mechanism layer. The feedforward neural network layer includes two linear transformation layers and a nonlinear activation function. The first linear transformation layer maps the input vector to a high-dimensional space, and then passes through the activation function to increase nonlinearity. The output is mapped back to the original space or the required target dimension through the second linear transformation layer.
[0059] The normalization layer is connected to the self-attention mechanism layer and the feedforward neural network layer respectively, and is used to normalize the outputs of the self-attention mechanism layer and the feedforward neural network layer to a preset numerical range, so as to accelerate the convergence speed of the model and improve the training stability;
[0060] The working principle and beneficial effects of the above technical solution are as follows: Figures 4 to 5 As shown in the figure, the BERT language model is used as the basis of the text error correction network model. A text error correction model suitable for error correction tasks is trained on this model. The BERT language model is actually a language processing tool. The Transformer model is added to the model, which includes a self-attention mechanism, a feedforward neural network, and a normalization method. This allows the trained model to have better language processing capabilities. The Transformer model is composed of an encoder and a decoder. The encoder extracts the features of the sentence and provides them to the decoder. Finally, the decoder completes the sentence operation through the features. Figure 4 As shown in the figure, the encoder understanding diagram in Transformer, the Encoder part of BERT is composed of multiple Transformer encoder layers stacked together, each Transformer encoder layer contains multiple self-attention heads, which can be calculated in parallel;
[0061] The BERT language model includes several encoders, including:
[0062] A self-attention layer is used to capture the dependencies between positions in the input sequence. By calculating the correlation between each position and all other positions, the model is able to generate updated representations for each position that contain information about other positions in the sequence, allowing the model to understand the context of the text.
[0063] The feedforward neural network layer is located after the self-attention mechanism layer. The feedforward neural network layer is responsible for further processing the output of the self-attention mechanism layer. The feedforward neural network layer usually contains two linear transformation layers and a nonlinear activation function (such as ReLU). The first linear transformation layer maps the input vector to a high-dimensional space, then applies the ReLU activation function to increase the nonlinearity, and finally maps the output back to the original space or the desired target dimension through the second linear transformation layer;
[0064] The normalization layer is located after the self-attention mechanism layer and the feedforward neural network layer. It is used to normalize the layer output to a certain numerical range (such as between 0 and 1) to accelerate the convergence speed of the model and improve the training stability.
[0065] like Figure 6 As shown, the calculation diagram of the self-attention score matrix of the self-attention mechanism in Transformer.
[0066] Q=XW Q
[0067] K=XW K
[0068] V=XW V
[0069] X is the feature extracted from the input sequence by the encoder, W Q , W K , W V Represents the weight matrix. For each position in the input sequence, a different weight matrix is used to linearly transform the input features to obtain the corresponding query vector Q, key vector K, and value vector V;
[0070] Then the normalized weight coefficient is calculated through the softmax function; QK T represents the dot product of the query vector Q and the key vector K; d K It is the dimension of the key value, which is used to scale the dot product to prevent the gradient from disappearing. The self-attention score matrix Attention(Q, K, V) is calculated by the weight coefficient and the value vector. The calculation formula is as follows:
[0071]
[0072] And each Transformer encoder layer has a feedforward neural network layer. The feedforward neural network layer consists of two fully connected layers (linear transformation layers) connected by a nonlinear activation ReLU function. This feedforward neural network layer can perform nonlinear transformation and mapping on the representation of each position output by the self-attention layer. The calculation formula is as follows
[0073] FFN(x)=max(0,xW1+b1)W2+b2
[0074] Among them, FFN(x) is the final calculation result of the feedforward neural network, x represents the self-attention score matrix, max is the ReLU activation function; W1 and W2 are both weight matrices; b1 and b2 are both bias terms;
[0075] The normalization layer maps the vector value to between 0 and 1. The output of the self-attention mechanism adds the input of the sub-layer to the output of the sub-layer through the residual connection, thereby retaining the information of the original input; layer normalization normalizes the result of the residual connection so that the input of each sub-layer has a similar mean and variance; to enhance information flow and alleviate the gradient vanishing problem, thereby accelerating the convergence of the model. The calculation formula is as follows:
[0076]
[0077] Among them, LayerNorm(x) represents the output value after rescaling and offset, which is used to restore data expression ability and enhance model flexibility. μ and σ are the mean and standard deviation of the input x, respectively. ∈ is a small constant used to prevent division by zero. γ and β are learnable scaling and offset parameters.
[0078] In one embodiment, the preprocessed corpus is input into the constructed BERT language model for training to obtain a BERT pre-trained language model, including:
[0079] A1: Initialize the model parameters of the built BERT language model;
[0080] A2: Input the first training set into the BERT speech model for iterative training. After the training is completed, a trained BERT language model is obtained.
[0081] A3: Input the first test set and the first valid set into the training BERT language model, test the training effect of the training BERT language model, obtain the test results, and optimize the model parameters of the BERT language model;
[0082] A4: Repeat steps A2 to A3 for a preset number of times to obtain a number of test results and a number of trained BERT language models, and screen the number of trained BERT language models based on the test results to obtain a BERT pre-trained language model;
[0083] The working principle and beneficial effects of the above technical solution are as follows: Figure 6 As shown, the first training set, the first test set and the first valid set divided from the preprocessed corpus are configured in the static folder of the model, and then some model parameters of the BERT language model are initialized; for example, the initial value settings of the model parameters include setting the learning rate to 0.70, the batch size to 16, the total number of iterations to 5, the optimizer to select the Adam optimizer, the weight decay to 0.02, the sequence length to 256, and the mask ratio model parameter to 0.20, thereby completing the initialization of the BERT language model; the first training set is input into the BERT language model for iterative training to obtain a trained BERT language model; the model training effect of the trained BERT language model is tested, and the test results are obtained by testing the first test set and the first valid set; at the same time, the model parameters are adjusted and training tests are performed to obtain several test results and several training BERTs. T language model; wherein, several test results correspond to several training BERT language models one by one; after screening according to the test results, the training BERT language model with the best test results is screened out as the BERT pre-training language model; for example, the initial model parameters are the learning rate set to 0.70, the batch size set to 16, the total number of iterations set to 5, the optimizer selected Adam optimizer, the weight decay set to 0.02, the sequence length set to 256, and the mask ratio model parameter set to 0.20; after multiple trainings, a set of model parameter data with the best training model effect is screened out, that is, the learning rate is set to 0.78, the batch size is 32, the total number of iterations is 11, the optimizer selects Adam optimizer, the weight decay is 0.02, the sequence length is 128, and the mask ratio model parameter is 0.15; at this point, the training phase of the BERT pre-training language model is completed.
[0084] In one embodiment, the BERT pre-trained language model is fine-tuned according to the pre-processed error correction task dataset to obtain a text error correction network model, including:
[0085] B1: Load the BERT pre-trained language model;
[0086] B2: Input the second training set into the BERT pre-trained language model for iterative training. After the training is completed, a training text error correction model is obtained.
[0087] B3: Input the second test set and the second valid set into the training text error correction model, evaluate the error correction performance of the training text error correction model, obtain the evaluation results, and fine-tune the model parameters of the training text error correction model;
[0088] B4: Repeat steps B2-B3 for a preset number of times to obtain a number of evaluation results and a number of training text error correction models, and screen the training text error correction models based on the evaluation results to obtain a training text error correction network model;
[0089] The working principle and beneficial effects of the above technical solution are as follows: Figure 7 As shown, the BERT pre-trained language model is loaded; the second training set divided by the error correction task data set is input into the BERT pre-trained language model for iterative training to obtain a training text error correction model; the error correction performance of the training text is evaluated by the first test set and the first valid set to obtain an evaluation result; at the same time, the model parameters are fine-tuned and fine-tuned and evaluated again to obtain several evaluation results and several training text error correction models; wherein, the several evaluation results and the several training text error correction models correspond one to one; after screening according to the evaluation results, the training text error correction model with the best test result is screened out as the text error correction network model; wherein, the evaluation of the training text error correction network model includes checking the evaluation indicators such as the accuracy, precision, recall rate and F1 score of the training results; for example, the error correction performance data is 0.6984 in accuracy, 0.6354 in recall rate and 0.6 in F1 score. 654, the system performance processing index is 58 times / s; such data results reflect that the text error correction performance is insufficient; by changing the learning rate, batch size and other models and conducting multiple trainings, new optimal performance index parameters are obtained; the optimal performance hyperparameters are learning rate set to 0.75, batch size set to 32, total number of iterations set to 10, optimizer selected Adam optimizer, weight decay set to 0.01, sequence length set to 128, mask ratio model parameter set to 0.15; after adjusting the parameters, the error correction accuracy rate increased from 0.6984 to 0.8254, the recall rate increased from 0.6354 to 0.7311, and the F1 score increased from 0.6654 to 0.7754, the system performance processing index increased from 58 times / s to 224 times / s, and the performance was greatly improved; the text error correction network model with the best error correction performance obtained through training was saved.
[0090] The present invention also provides a text comparison system, comprising: a text acquisition module, used to obtain one or more error correction texts to be processed;
[0091] The text correction module is used to process one or more error correction texts to be processed using a pre-trained text correction network model to obtain the correct text after correction; wherein the text correction network model is obtained by the above-mentioned training method of the text correction network model based on deep learning.
[0092] The present invention also provides a computer device, comprising:
[0093] Memory for storing computer programs;
[0094] A processor is used to implement the training method of the text error correction network model based on deep learning as described in any of the above items when executing the computer program.
[0095] The present invention also provides a computer-readable storage medium, on which a computer program is stored, and the computer program is executed by a processor to implement the training method of the text error correction network model based on deep learning as described in any one of the above.
[0096] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solution of the present invention, which should be included in the scope of the claims of the present invention.
Claims
1. A training method for a text error correction network model based on deep learning, characterized in that: include: Obtain training error correction documents and several corpora; Preprocess the training error correction documents and several corpora to obtain the preprocessed error correction task dataset and preprocessed corpus respectively; Input the preprocessed corpus into the constructed BERT language model for training to obtain the BERT pre-trained language model; The BERT pre-trained language model is fine-tuned according to the pre-processed error correction task dataset to obtain a text error correction network model.
2. A training method for a text error correction network model based on deep learning as claimed in claim 1, characterized in that: Preprocess the training error correction documents and several corpora to obtain the preprocessed error correction task dataset and preprocessed corpus, including: Processing the obtained corpora according to a first preset processing strategy to obtain a preprocessed corpus; wherein the preprocessed corpus includes: a first training set, a first test set and a first valid set; The acquired training error correction document is processed according to a second preset processing strategy to obtain a preprocessing error correction task data set; wherein the preprocessing error correction task data set includes: a second training set, a second test set and a second valid set.
3. The method for training a text error correction network model based on deep learning as claimed in claim 1, characterized in that: The constructed BERT language model includes several encoders and several encoder stacking settings. The encoders include: A self-attention mechanism layer to capture the dependencies between positions in the input sequence; The feedforward neural network layer is connected to the self-attention mechanism layer and is used to further process the output of the self-attention mechanism layer. The feedforward neural network layer includes two linear transformation layers and a nonlinear activation function. The first linear transformation layer maps the input vector to a high-dimensional space, and then passes through the activation function to increase nonlinearity. The output is mapped back to the original space or the required target dimension through the second linear transformation layer. The normalization layer is connected to the self-attention mechanism layer and the feedforward neural network layer respectively, and is used to normalize the outputs of the self-attention mechanism layer and the feedforward neural network layer to a preset numerical range to accelerate the convergence speed of the model and improve the training stability.
4. A method for training a text error correction network model based on deep learning as claimed in claim 2, characterized in that: Input the preprocessed corpus into the constructed BERT language model for training to obtain the BERT pre-trained language model, including: Initialize the model parameters of the built BERT language model; The first training set is input into the BERT speech model for iterative training, and after the training is completed, a training BERT language model is obtained; Inputting the first test set and the first valid set into the training BERT language model, testing the training effect of the training BERT language model, obtaining the test results, and optimizing the model parameters of the BERT language model; Repeat the above steps to obtain several test results and several trained BERT language models, and screen the several trained BERT language models based on the test results to obtain a BERT pre-trained language model.
5. The method for training a text error correction network model based on deep learning as claimed in claim 2, characterized in that: The BERT pre-trained language model is fine-tuned according to the pre-processed error correction task dataset to obtain a text error correction network model, including: Load the BERT pre-trained language model; The second training set is input into the BERT pre-trained language model for iterative training, and after the training is completed, a training text error correction model is obtained; Inputting the second test set and the second valid set into the training text error correction model, evaluating the error correction performance of the training text error correction model, obtaining an evaluation result, and fine-tuning the model parameters of the training text error correction model; Repeat the above steps to obtain several evaluation results and several training text error correction models, and screen the several training text error correction models based on the evaluation results to obtain a training text error correction network model.
6. A text comparison system, characterized in that: include: A text collection module, used to obtain one or more error correction texts to be processed; A text correction module is used to process one or more error correction texts to be processed using a pre-trained text correction network model to obtain a correct text after correction; wherein the text correction network model is obtained by the training method of the text correction network model based on deep learning according to any one of claims 1 to 5.
7. A computer device, characterized in that: include: Memory for storing computer programs; A processor, configured to implement the training method for a text error correction network model based on deep learning as described in any one of claims 1 to 5 when executing the computer program.
8. A readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and the computer program is executed by a processor to implement the training method of a text error correction network model based on deep learning as described in any one of claims 1 to 5.
Citation Information
Patent Citations
BERT-based Chinese ASR output text repairing method and system
CN115034208A
Model training method, Chinese text error correction method, electronic equipment and storage medium
CN116167362A
Medical natural language processing method based on pre-training model
CN116341546A
Geographic named entity error correction method based on pre-training deep learning
CN116702760A
